U-Net-based Transmission Line Segmentation Network
Segmentation of transmission line images has low resolution and accuracy owing to complex line background,interweaving of lines and blurred targets on towers.In view of this the present work carried out a study on U-Net-based segmentation networks.First VGG16,ResNet34,ResNet50,MobileNetV2 and DenseNet were used as the backbone fea-ture extraction networks of U-Net networks.Second BCEWithLogitsLoss and DiceLoss were used as network loss func-tions.Finally a unified data set was used to study the segmentation utility of several loss functions and network structures by analyzing their image segmentation accuracies.Experimental results showed that under this specific segmentation sce-nario,Dense Net and VGG16 models with BCEWithLogitsLoss and DiceLoss mixed loss functions have better line seg-mentation utilities.
U-Nettransmission lineimage segmentationloss function
NETL
NSTL
万方数据
维普
